Global Memory — DRAM · NAND · HBM, with China in the Frame (Memory China Sector)

Executive Summary & Action Plan

Verdict

SELL / UNDERWEIGHT — Every name in the complex trades above fair value once peak margins are normalized, and none earns ownership at these levels. The honest anchor is Samsung Electronics (005930.KS), the only one whose reverse-DCF is arithmetically reachable: spot demands a 31.4% blended terminal operating margin, against 90.8% for SK hynix and 94.0% for Micron. China is not the reason to sell today; it is the reason the eventual normalization will not be followed by the usual oligopoly recovery.

Price ₩244,000 | Market cap ₩1,406.5T | Target ₩202,000 | Upside −17.2% IWANNAVY Fair Value (peak-margin-normalized) · price as of 2026-07-21 close · Street consensus ₩492,537 (n=36, reference) — we sit 58.9% below Street because consensus applies a mid-cycle multiple to near-peak memory earnings, where we apply it to mid-cycle earnings and credit the boom only through an explicit discounted cycle-surplus bridge

Abstract

We underwrote the global memory complex at 2026-07-21 closes with full cyclical discipline and extended coverage beyond Korea to Micron, Taiwan and the unlisted Chinese makers. The earnings are the highest in the industry's history — Micron at an 80.4% operating margin and SK hynix at 71.5%, roughly 25–30 points above any prior-cycle record — and the decisive question is whether that is earnings or a price spike. It is a price spike: gross margins near 85% imply ASPs running five to six times cash cost, and Nanya prints 73.7% on pure commodity DRAM having printed −46.5% in Q4'23. Normalizing margin and revenue together yields Samsung ₩202,000, SK hynix ₩1,170,000 and Micron $500, leaving the group 17% to 42% above fair value. The bifurcation is real: China holds roughly 10% of DRAM bits and 13% of NAND revenue but near zero merchant HBM. Ratings are SELL across; the 7/29 SK hynix print and the 7/27 CXMT debut are the near tests.

Action Plan

Reduce into strength rather than chase the bounce; re-entry belongs at normalized fair value, and Samsung is the only name whose ladder is worth staging.

The ranking below is the whole report in one frame: upside to our normalized fair value, not to Street targets.

• Prob-weighted 12M return: Samsung +1.3% (Bear 26 / Base 46 / Bull 28) · SK hynix −26.0% (32/44/24) · Micron −37.6% (36/46/18) | p_thesis_wrong 0.62 | Confidence: Medium-High on normalization, Medium on cycle timing — long-term agreements can extend a peak past what history suggests.

---

What Happened / State of the Sector

Monday's session was the tell. Seoul opened on record data — Korean semiconductor exports of $22.1bn for July 1–20, up 180.6% year over year, an all-time record for the period — and SK hynix was up 3.0% by 11:00 KST, Samsung up 5.1%. Both closed down more than 4%, a seven-to-nine point round trip in one session. A market that sells a +180% export print is not trading demand; it is trading supply of stock.

Three forces explain it. First, Korea's forced-deleveraging machine: margin loans peaked at ₩38.63T and roughly 1.2 million accounts hit call thresholds in July, with 320,000–360,000 liquidated, prompting emergency FSC measures on 7/16. Second, and most damaging, SK Group chairman Chey Tae-won publicly called memory prices "abnormally high" and argued they "have to normalize — otherwise the market shrinks and competitors flood in," pairing it with accelerated Yongin capacity. That is the controlling shareholder of the largest HBM supplier repudiating his own run-rate margins, 24–48 hours before the close. Third, CXMT priced the largest A-share chip IPO in history on 7/14 — RMB 57.9bn (~$8.55bn) at a ~$85bn valuation, institutional book 462.85x oversubscribed — with its trading debut on 7/27. The damage ranked exactly as that implies: commodity pure-play Nanya (2408.TW) fell hardest at −3.67%, niche Winbond (2344.TW) −2.58%.

The composition of the boom is the second underappreciated fact. HBM is the cause of the supercycle but commodity DRAM is where the profit landed. HBM consumes about three times the wafer area per bit and took ~23% of DRAM capacity in 2026, manufacturing the shortage — yet it is sold on three-to-five-year agreements that capped its own pricing while conventional DDR5 contracts rose 90–95% in a single quarter. That is why SK hynix, the supplier with the most AI exposure, lost 4.1 points of DRAM revenue share in Q1'26, and why it delayed converting HBM3E lines to HBM4 in June to build DDR5 instead.

Where the money is made, and how much of it China can reach, is the sector's central geography. The two charts below pair DRAM against NAND deliberately.

Margins across all three majors trace the same blow-off, from different starting points. Samsung's lower line is dilution from foundry, mobile and display, not weaker memory.

SK hynix's own income flow shows how violent the operating leverage is in both directions: revenue tripled year over year while cost of goods rose only 45%.

---

China: Winning One Industry, Locked Out of the Other

Memory has split into two businesses that share a name and almost nothing else, and China's position in each is opposite.

In commodity bits, China is already first-tier. CXMT is the world's #4 DRAM maker at roughly 8% of revenue and 10–11% of bits, running about 265k wafer starts per month mid-2026 and heading to ~350k by December — which matches Micron's ~375–385k scale. Its LPDDR5X-10667 meets flagship mobile spec, LPDDR is over 40% of its shipments and designed into more than 30% of Android handsets sold in China. YMTC is at 13% of NAND revenue, up from 8% a year ago — the fastest share gain of any supplier — shipping 294-layer product against Samsung's 286L. NAND is the exception that explains everything: 3D NAND scales by stacking and etching, not by lithographic resolution, so EUV denial costs YMTC almost nothing. It invented the Xtacking hybrid-bonding architecture and Samsung licensed those patents from YMTC for its V10-and-beyond NAND, the first time a Chinese memory vendor reverse-licensed a global leader.

In AI memory, China's share is approximately zero and will stay there. CXMT's HBM3 mass production was targeted for the first half of 2026 and has slipped, judged unlikely within the year and still in test, against Korean HBM3 mass production that began in 2023. Overall HBM yield is modelled near 25%. China-wide domestic HBM output for 2026 is roughly 2 million stacks, captive to Huawei Ascend, against a merchant market measured in tens of millions. We assess the functional merchant-HBM gap at 4–6 years and widening, above the 3–4 years commonly cited, because four bottlenecks are sequential and independent rather than overlapping: HBM3 mass production not yet reached; 12-hi and 16-hi stacking tools controlled since December 2024; an HBM4 base die that requires a leading-edge logic node SMIC cannot supply at any capex level; and HBM4E-class core die requiring 1c DRAM, which requires EUV that has never shipped to China.

The real China risk is not the one being debated. It is not that China takes HBM. It is that China takes enough of conventional DRAM that the non-HBM half of the P&L reverts to trough economics while the whole company is still priced as if it were HBM. CXMT is phasing out server and PC DDR4 by mid-2026 and redeploying that capacity into DDR5, adding roughly 85k wafers in 2026 and 70k in 2027 — from a producer whose cost per bit is more than 30% above the incumbents while its ASP is only 5–10% below. Near-parity pricing on a 30% cost handicap survives only in shortage. And the classic discipline mechanism, marginal producers going bankrupt, has been switched off: Hefei state capital owns ~36.8% of CXMT, the Big Fund holds 8.73%, and the July IPO added $8.5bn of public money. Subsidy buys capacity, not yield — but it also buys survival, and that permanently lowers through-cycle memory returns. Add Fujian Jinhua's 60k wpm phase one and SwaySure's 140k wpm design and China's DRAM wafer base approaches 500–550k wpm by 2027–28 (est.). Almost none of it is HBM.

One regime change deserves flagging because most China-threat commentary has missed it: the 2024–25 Chinese DDR4 price war is over. Chinese pricing now sits near Korean levels, Beijing ordered DDR4 output halted and curbed memory speculation. China is no longer competing on price but on allocation, mandate and capacity — a more durable and more dangerous form of competition.

---

What to Watch Now

• SK hynix Q2'26 results, 7/29 (D-8): consensus revenue ₩84.7T and operating profit ₩64.9T, but the working band is ₩60.4T–₩70T. Q2 was still a price-capped quarter — SK hynix only removed LTA ceilings in July — so spot strength does not pass through. Guidance on 2H26 supply and LTA structure matters more than the print.
• CXMT trading debut, 7/27: the first public market-clearing price for Chinese DRAM. A strong debut hands a commodity competitor a permanent, low-cost war chest.
• 4Q26 contract negotiations, September–October: the first genuine price-discovery round of the post-sold-out era. The first quarter guided flat or down is the signal.
• Micron FQ4, ~late September (D-70): a positioning question, not an event trade. Micron's LTAs retain price caps set at Q2'26 levels, so calendar Q3 strength largely cannot pass through. The scenario to fear is an in-line FQ4 with FQ1 FY27 guided below $50B, which would date the peak.
• Customer inventory: 7–9 weeks, up from 2–4 weeks in February 2026. Crossing 10 weeks is the mechanical trigger.

What to Watch Next (Technology)

• 2027 HBM4 contract ASPs — the thesis breaker. If they settle flat or down versus 2026 HBM3E, "durably higher margins" is falsified at its source.
• HBM qualification has already de-moated. On 5 June 2026 NVIDIA confirmed all three suppliers qualified and in production for Vera Rubin HBM4, ending SK hynix's sole-source era. Q2'26 HBM share: SK hynix 62%, Micron 21% — having overtaken Samsung, the most consequential share move of the year — Samsung 17%, down from 21%.
• CXMT–YMTC hybrid-bonding alliance: CXMT supplying DRAM, YMTC supplying Xtacking, targeting a 2027–28 domestic HBM3E window for Huawei Ascend. HBM is the binding constraint on China's entire AI accelerator programme — SMIC can produce die for over a million Ascend chips a year, but domestic HBM caps output below 300,000.
• The MATCH Act and a possible CXMT Entity Listing. CXMT is not Entity-Listed, uniquely among China's four largest fabricators — the asymmetry that let it scale from 40k to 265k wspm. A listing would be a step-function derating of its roadmap. But 2026 policy has moved the other way: the Affiliates Rule is suspended for a year under the trade truce.

Fair Value Notes

• Samsung Electronics (005930.KS) ₩202,000 — log-median of DCF ₩191,718 (WACC 10.3%, g 2.5%, normalized blended operating margin 21%), 5.0x mid-cycle EBITDA ₩188.6T → ₩202,441, and 13x mid-cycle FCF ₩81.9T → ₩223,502. Its TTM EBITDA margin is already near mid-cycle because non-memory dilutes the spike, so its 9.15x EV/EBITDA is the least distorted starting point in the group. It converts a 4.80% FCF yield at the peak, roughly 2.3x SK hynix's cushion and 6x Micron's. Supersedes our 7/13 mark of ₩150,000: same normalization discipline, plus an explicit discounted cycle-surplus bridge (₩104.9T) that credits the largely contracted windfall the balance sheet will genuinely absorb.
• SK hynix (000660.KS) ₩1,170,000 — log-median of DCF ₩912,548 (WACC 11.0% on Korea 10Y 4.33%, normalized operating margin 35%), 6.5x mid-cycle EBITDA ₩113.2T → ₩1,171,328, 16x mid-cycle FCF ₩53.5T → ₩1,340,202. Bull ₩2,650,000, bear ₩484,000. Reverse-DCF: spot demands ₩611T of 2031 revenue at a 35% margin, or a 90.8% terminal operating margin on our base — arithmetically unattainable. Supersedes 7/13's ₩1,060,000 on the same cycle-surplus addition.
• Micron (MU) $500 — log-median of DCF $400.95 (WACC 11.1% on US 10Y 4.60%, normalized operating margin 32%), 6.5x mid-cycle EBITDA $73.2B → $500.67, 16x mid-cycle FCF $36.0B → $589.94. Bull $1,298, bear $212. Worst risk-reward of the three: a 0.78% FCF yield and 125.4x EV/FCF mean almost none of the windfall is reaching shareholders before the cycle turns, and its LTAs retain caps that truncate upside. This supersedes our stale 6/30 mark of $1,300, which was set under the weaker normalization framework the house abandoned on 7/13; carrying it forward would have produced a spurious "sell Korea, buy Micron" conclusion driven by methodology rather than economics.
• Taiwan, screened not modelled: Nanya (2408.TW) looks cheap at 9.65x EV/EBITDA only because the denominator is peak — a 73.7% operating margin against a through-cycle average near zero implies roughly 25–40x normalized, the most peak-earnings risk in the group. Winbond (2344.TW) is worse at 23.2x on inflated earnings with a negative FCF yield, near 40–55x normalized.

Our fair value work rests on three independent methods; the dot plot shows where each lands for the anchor.

• Thesis breaker: 2027 HBM4 contract ASPs settling flat or down versus 2026 HBM3E, as reported on SK hynix's late-October call.

---

Catalysts & Risks

---

References

• Korea Herald — July 1–20 exports +52%, semiconductors $22.1bn +180.6% (2026-07-21) · Korea Herald — Chey Tae-won calls memory prices abnormally high (2026-07) · Tom's Hardware — SK Group chairman on "abnormally high" RAM prices
• Bloomberg — CXMT prices Shanghai STAR IPO at RMB 8.66 (2026-07-14) · SemiAnalysis — China's CXMT set to challenge DRAM incumbents · Tom's Hardware — CXMT close to matching Micron capacity in 2026
• TrendForce — 3Q26 memory price gains moderate (2026-07-03) · TrendForce — LTAs cap server DRAM price increases (2026-07-09) · Tom's Hardware — memory price surge begins to cool
• Counterpoint — global DRAM revenue near $100B in Q1'26 · Counterpoint — NAND record $46B in Q1'26 · TrendForce — top-five NAND revenue +83.7% QoQ in 1Q26 (2026-05-25)
• CNBC — SK hynix worst day, −15%, after Nasdaq debut (2026-07-13) · KED Global — Korea tightens single-stock leveraged ETF rules (2026-07-16) · DigiTimes — shortage could turn to glut by 2028 (2026-07-20)